Releases Index
OpenRouter

OpenRouter

$npx @buildinternet/releases get openrouter
Recently Shipped30 releases · updated Nov 17, 2026

Recently shipped new Auto and Auto Code routers, a rebuilt provider routing algorithm, and an alpha Extractions API that pulls typed fields out of text, images, and PDFs.

The default router now weighs speed, throughput, and reliability alongside price, not price alone. Providers that deliver better inference win more traffic automatically, and no code changes are needed to benefit.

  • The rebuilt Auto Router (openrouter/auto) beats the old one by up to 4× depending on the benchmark, with a 3× faster routing step at p90 (0.65s vs 1.96s).
  • Auto Code (openrouter/auto-code) is a new router tuned for coding work.
  • Across live traffic, the new routing lifted throughput 16.3%, cut time to first token 16.9%, and reduced capacity errors 67% for a 6% cost increase.

Guardrails now separate organization-wide limits from per-person exceptions. Admins set one set of rules for the whole workspace, then give individuals different budgets and model access within them. Each policy is edited on its own.

Private endpoints can now be updated with a single merge patch. PATCH /private-endpoints/{id} changes only the fields you send, so no field is required. Two action items follow from the endpoint cleanup. Calls to PUT /private-endpoints/{id}/parameters and PUT /private-endpoints/{id}/pricing now return 404; send declared_parameters and pricing to the PATCH route instead. Private providers no longer accept a data_policy field, so drop it from POST /private-providers calls and Terraform configs.

S3 observability destinations must point at an IAM role in the commercial AWS partition with an sts:ExternalId trust condition, or test and live delivery fail with 403. Python SDK openrouter 1.3.70+ and TypeScript @openrouter/sdk 1.5.0+ renamed the EndUser models to ExternalUser with unchanged shapes.

Hugging Face inference runs on OpenRouter routing

This release3 featuresNew capabilitiesAI-tallied from the release notes

Chat models on Hugging Face Inference Providers now route through OpenRouter, which picks providers by live price, speed, and reliability and fails over automatically when one slows or errors. OpenRouter can also broadcast each request trace to a Hugging Face Storage Bucket you own, alongside a new Providers tab with live price, speed, and uptime per provider.

Read more ↗

Management keys now scoped to one workspace; read-only option added

This release1 featureNew capabilities1 enhancementImprovements to existing featuresAI-tallied from the release notes

Management keys can now be limited to a single workspace and set to read-only, so a key can view API keys, activity, and usage without gaining write access or reaching other workspaces. Scope and access level are fixed at creation, and budget limits, credits, and organization settings still require an organization key.

Read more ↗

Extractions API ships; unknown field keys now rejected

This release3 featuresNew capabilities5 enhancementsImprovements to existing featuresAI-tallied from the release notes

POST /api/alpha/extractions is now available, and extraction request and field schemas reject unknown keys instead of silently dropping them, returning 400 with a validation error naming the key. Also removed the no-op provider.options.scaledown.context_chars, added a per-item confidences array on list answers, PDF asset support, and new filter schemas for the analytics query endpoint.

Read more ↗

Deferred tools gain cascade mode and hint properties

This release4 featuresNew capabilities5 enhancementsImprovements to existing featuresAI-tallied from the release notes

DeferredToolsControl adds a cascade_mode property and a hint property, and DeferredBm25Search and DeferredRegexSearch gain a siblings property. Several tool schemas, including ImageGenerationServerToolConfig and CodeInterpreterServerTool, received updated descriptions and examples.

Read more ↗

MCP tool events for Responses API; deferred tools in usage

This release9 featuresNew capabilities6 enhancementsImprovements to existing featuresAI-tallied from the release notes

OpenRouter's Responses API now reports MCP tool calls and tool listing through new streaming and output schemas, with tool choice and approval requests included. ChatUsage gains a deferred_tools_details field, and decisions requests no longer require the instructions property.

Read more ↗

Machine credit purchases; extraction spans can be null

This release5 featuresNew capabilities10 enhancementsImprovements to existing featuresAI-tallied from the release notes

New endpoints let machines buy credits and check purchases, presets gain restore and workspace_id filtering, and extraction responses now return 200 with null entries in spans instead of 502 when list items have no offsets in the state text. Client code reading spans must check each entry for null.

Read more ↗

Private-endpoint PUT routes removed; provider endpoints list added

This release3 featuresNew capabilities10 enhancementsImprovements to existing featuresAI-tallied from the release notes

The deprecated PUT /private-endpoints/{id}/parameters and /pricing endpoints now return 404, and POST /private-endpoints no longer reads a top-level max_completion_tokens field, which must move to settings.max_completion_tokens. GET /providers/{provider_slug}/endpoints lists a provider's endpoints with provider-wide stats, and STT, speech, and video requests accept provider routing fields.

Read more ↗

Deferred search and tool-control schemas gain new properties

This release6 featuresNew capabilities3 enhancementsImprovements to existing featuresAI-tallied from the release notes

DeferredBm25Search and DeferredRegexSearch gained spread and miss_sources properties, while DeferredToolsControl adds announce_catalog_changes and min_deferred_tools. OutputToolSearchServerToolItem also gained a catalog_changes property, and several schema descriptions were updated.

Read more ↗

OTEL destinations get metrics config; guardrail schema refs updated

This release5 featuresNew capabilities5 enhancementsImprovements to existing featuresAI-tallied from the release notes

Observability OTel collector destinations gain config.metricsEndpoint and config.shouldSendMetrics properties, and web fetch tool output now includes a retrieved_at timestamp. The changelog also adds a ContentFilterReadEntry guardrail schema, bad-request response schemas for POST /responses and credits purchases, and updates the embeddings request and response schemas.

Read more ↗

Models API gains endpoints parameter

This release3 featuresNew capabilities4 enhancementsImprovements to existing featuresAI-tallied from the release notes

The GET /models endpoint accepts a new endpoints parameter. The API also adds Pearl Research and Ant Digital provider options, enum additions for reasoning format and BYOK provider slugs, and an input_tokens_details property on decisions usage.

Read more ↗

deferred_tools "client" scope removed; "all" and "mcp" remain

This release1 featureNew capabilities3 enhancementsImprovements to existing featuresAI-tallied from the release notes

Sending "scope": "client" in deferred_tools now returns 400 on chat/completions, messages, responses, and preset endpoints; scope accepts only "all" or "mcp". Omit scope or set "all" to defer across all eligible tools, or set "mcp" to defer MCP tools only.

Read more ↗

Extractions and guardrail schemas drop required workspace_id

This release3 featuresNew capabilities15 enhancementsImprovements to existing featuresAI-tallied from the release notes

ExtractionsRequest no longer requires a state property, and the guardrail policy request schemas no longer require workspace_id, with the changelog noting no action is needed. ChatRequest also gains reasoning.context, reasoning.enabled, reasoning.exclude and reasoning.mode properties, and extraction field schemas add unit and integer properties.

Read more ↗

Guardrail policies get full CRUD API; presets accept workspace_id

This release23 featuresNew capabilities4 enhancementsImprovements to existing featuresAI-tallied from the release notes

New endpoints let workspaces create, list, update, and delete account-wide guardrail policies covering budget, moderation, and model availability, plus read and change policy assignments for API keys and members. Preset chat, messages, and responses endpoints accept a new workspace_id parameter, and DeferredToolsControl gained interface and scope properties.

Read more ↗

Private providers drop configurable data_policy

This release5 enhancementsImprovements to existing featuresAI-tallied from the release notes

Private provider requests and responses no longer accept or return a configurable data_policy, and newly created providers automatically apply a policy with no prompt retention, training, or publication. A PATCH body containing only data_policy now returns 400 instead of 200.

Read more ↗

OpenRouter's site is now available in Simplified Chinese, Spanish, Brazilian Portuguese, German, French, Japanese, and Korean alongside English, in beta, selectable from the account menu or via locale URLs like openrouter.ai/ja. Translations live as JSON catalogs in Git and run through a daily pipeline that extracts changed strings by fingerprint, translates them, checks placeholders and glossary terms in code, and has a verifier model flag errors before an AI reviewer and a human merge the pull request.

Read more ↗

Extractions API extracts typed fields from text, images, PDFs

This release3 featuresNew capabilitiesAI-tallied from the release notes

OpenRouter's Extractions API is live in alpha, returning one typed value per named field from text, images, and PDFs across nine field types, with spans, p_null, and confidence on supported models. It launches with Cloudflare Clairic for text and images, ScaleDown Extract for PDFs, and Inception Mercury Decide for text only.

Read more ↗

Auto Router rebuilt; Auto Code tuned for coding harnesses

This release2 featuresNew capabilities1 enhancementImprovements to existing featuresAI-tallied from the release notes

OpenRouter rebuilt Auto Router on a new foundation that weighs eleven factors per request alongside live pricing, provider health, and cache status, with routing 3x faster at p90. A new Auto Code router (openrouter/auto-code) targets coding harnesses and scores in line with Claude Opus 5.5 at 74% less cost per task, and the default cost tier is now closer to pricing from a few months ago.

Read more ↗

Guardrail limits now set by org, defaults by workspace

This release3 featuresNew capabilities1 enhancementImprovements to existing featuresAI-tallied from the release notes

Organization admins can now set workspace-wide limits on budget, Zero Data Retention, and model and provider access that apply to every member and API key, while workspace admins set defaults inside those limits. Guardrails also split into separate budget, model access, and content moderation policies, and existing guardrails keep their IDs and rules.

Read more ↗

Private providers API ships; PATCH is now merge patch

This release6 featuresNew capabilities3 enhancementsImprovements to existing featuresAI-tallied from the release notes

New /private-providers CRUD endpoints manage private providers, and PATCH /private-endpoints/{id} is now a merge patch over every owner-controlled field, so upstream_model_id is no longer required. The pricing and parameters PUT routes remain as deprecated aliases.

Read more ↗